SOCP转SDP的几何直观及旋转锥等价关系的直觉理解问询

Great question—algebraic proofs get you the "what" of this equivalence, but the geometric intuition is what makes it click for why we use this transformation in SOCP. Let's break this down step by step:

1. Start with the Standard Second-Order Cone (Lorentz Cone) #

First, recall the standard Lorentz cone (the core of SOCP), which is a convex "ice cream cone" in $\mathbb{R}^{n+1}$:
$$\mathcal{K}_{n+1} = \left{ \begin{bmatrix} u \ t \end{bmatrix} \in \mathbb{R}^n \times \mathbb{R} \mid |u|_2 \leq t, , t \geq 0 \right}$$
Geometrically, every point in this cone lies within a distance $t$ from the $t$-axis (the cone's central axis), with $t$ being the non-negative "height" along that axis.

2. Rewrite the Quadratic Inequality as a Cone Constraint #

Your target quadratic inequality is:
$$x^T A^T A x + b^Tx + c \leq 0 \tag{2}$$
Notice that $x^T A^T A x = |Ax|_2^2$, so this is a convex quadratic constraint (since $A^TA$ is positive semi-definite). The trick is to map this into the Lorentz cone's structure using an affine transformation.

Let's define $s = b^Tx + c$ for simplicity. The left-hand side cone constraint becomes:
$$\left| \begin{bmatrix} Ax \ \frac{1+s}{2} \end{bmatrix} \right|_2 \leq \frac{1-s}{2} \tag{3}$$
First, note the implicit condition here: the right-hand side $\frac{1-s}{2}$ must be non-negative (since the left-hand side is a $L_2$ norm, which is always non-negative). This gives $1-s \geq 0 \implies s \leq 1$, or $b^Tx + c \leq 1$—a necessary condition for the constraint to have any feasible solutions.

3. Geometric Interpretation of the Transformation #

Think of this as a shifted and scaled section of the standard Lorentz cone:

• Let $t = \frac{1-s}{2}$ and $v = \frac{1+s}{2}$. The constraint (3) becomes $\left| \begin{bmatrix} Ax \ v \end{bmatrix} \right|_2 \leq t$.
• Notice that $t + v = \frac{1-s}{2} + \frac{1+s}{2} = 1$, so $v = 1 - t$. This means we're looking at the intersection of the standard Lorentz cone with the plane $v + t = 1$ in the $(Ax, v, t)$ space.

When we project this intersection back to the original $x$ space, we get exactly the region defined by the quadratic inequality (2). Geometrically, the quadratic inequality describes an ellipsoid (or a half-space/empty set if $A^TA$ is singular), and the cone constraint is just a way to represent this ellipsoid using the SOCP's native cone structure—since SOCP solvers are optimized to handle these cone constraints efficiently.

4. Why This Specific Transformation? #

The choice of splitting $1$ into $\frac{1+s}{2}$ and $\frac{1-s}{2}$ isn't arbitrary. When you square both sides of (3), the cross terms from the linear term $s$ cancel out perfectly, leaving you with the original quadratic inequality. Intuitively, this split lets us "package" the quadratic term $|Ax|_2^2$ and the linear term $s$ into the $L_2$ norm of a vector, which fits exactly into the Lorentz cone's definition of "distance from the axis ≤ height".

Key Takeaway #

This equivalence is all about converting convex quadratic constraints into standard SOCP cone constraints via affine transformations. Geometrically, you're taking the region defined by the quadratic inequality (an ellipsoid-like shape) and mapping it to a section of the Lorentz cone—something SOCP solvers are designed to handle natively.

内容的提问来源于stack exchange，提问作者jjjjjj